Automatic Optical Coherence Tomography Imaging Analysis for Retinal Disease Screening

(1)

Automatic Optical Coherence Tomography Imaging Analysis for

Retinal Disease Screening

Md Akter Hussain

Submitted in total fulfilment of the requirements of the degree of

Doctor of Philosophy

School of Computing and Information Systems THEUNIVERSITY OFMELBOURNE

August 2017

(2)

(3)

Abstract

The retina and the choroid are two important structures of the eye and on which the quality of eye sight depends. They have many tissue layers which are very important for monitoring the health and the progression of the eye disease from an early stage. These layers can be visualised using Optical Coherence Tomography (OCT) imaging. The abnormalities in these layers are indications of several eye diseases that can lead to blindness, such as Diabetic Macular Edema (DME), Age-related Macular Degeneration (AMD) and Glaucoma. If the retina and the choroid are damaged there is little chance to recover normal sight. Moreover, any damage in them will lead to blindness if no or late treatment is administered. With eye diseases, early detection and treatment are more effective and cheaper. Biomarkers extracted from these tissue layers, such as changes in thickness of the layers, will note the presence of abnormalities called pathologies such as drusen and hyper-reflective intra-retinal spots, and are very effective in the early detection and monitoring the progression of eye disease. Large scale and reliable biomarker extraction by manual grading for early detection is infeasible and prone to error due to subjective bias and are also cost ineffective. Automatic biomarker extraction is the best solution.

However, OCT image analysis for extracting biomarkers is very challenging because of noisy images, low contrast, extremely thin retinal layers, the presence of pathologies and complex anatomical structures such as the optic disc and macula. In this thesis, a robust, efficient and accurate automated 3D segmentation algorithm for OCT images is proposed for the retinal tissue layers and the choroid, thus overcoming those challenges. By map- ping OCT image segmentation problem as a graph problem, we converted the detection of layer boundaries to the problem of finding the shortest paths in the mapped graph.

The proposed method exploits layer-oriented small regions of interest, edge pixels from iii

(4)

shortest path algorithm as a boundary of the layers. Using this segmentation scheme, we were able to segment all the retinal and choroid tissue layers very accurately and extract eight novel biomarkers such as attenuation of the retinal nerve fibre layer, relative intensity of the ellipsoid zone, thickness of the retinal layers, and volume of pathologies i.e.

drusen, etc. In addition, we demonstrated that using these biomarkers provides a very accurate (98%) classification model for classifying eye patients into those with normal, DME and AMD diseases which can be built using a Random Forest classifier.

The proposed segmentation method and classification method have been evaluated on several datasets collected locally at the Center for Eye Research Australia and from the public domain. In total, the dataset contains 56 patients for the evaluation of the segmentation algorithms and 72 patients for the classification model. The method developed from this study has shown high accuracy for all layers of the retina and the choroid over eight state-of-the-art methods. The root means square error between manually delineated and automatically segmented boundaries is as low as 0.01 pixels. The quantification of biomarkers has also shown a low margin of error from the manually quantified values. Furthermore, the classification model has shown more than 98% accuracy, which outperformed four state-of-the-art methods with an area under the receiver operating characteristic curve (AUC) of 0.99. The classification model can also be used in the early detection of diseases which allows significant prevention of blindness as well as providing a score/index for the condition or prediction of the eye diseases. In this thesis, we have also developed a fully automated prototype system, OCTInspector, for OCT image analysis using these proposed algorithms and methods.

iv

(5)

Declaration

This is to certify that

1. the thesis comprises only my original work towards the PhD degree, 2. due acknowledgement has been made in the text to all other material used,

3. the thesis is less than 100,000 words in length, exclusive of tables, maps, bibliogra- phies and appendices.

Md Akter Hussain, August 2017

v

(6)

(7)

Acknowledgements

I praise Allah (the Creator and the Sustainer of the Worlds), the Most Gracious, the Most Merciful, for blessing me with the opportunity, courage, and intellect to undertake this research.

I would like to express my gratitude to my supervisor, Prof. Ramamohanarao (Rao) Kotagiri, for his patient guidance, constant support, helpful criticisms, valuable sugges- tions and commendable support for the completion of my thesis. I am profoundly in- debted to my co-supervisor Dr. Alauddin Bhuiyan for his constant guidance and insight- ful advice as an experienced researcher in the field of imaging. They have given me the freedom to explore research challenges of my choice and guided me when I felt lost. I am grateful to Prof. Robyn Guymer and Prof. Chi D. Luu for their guidance relating to eye research and for providing me with access to the data from the Center for Eye Re- search Australia (CERA). I would also like to thank Prof. Andrew Turpin for his support and encouragement and his advice as the chair of my Advisory Committee. I am also thankful to Prof. Smith, R. Theodore and Prof. Hiroshi Ishikawa, New York University, for their guidance and provision of data. I would like to thank Juan Sepulveda and Dr.

Fumi Tanabe for helping me in reviewing the manual segmentation of OCT images. I would like to thank Gregory Rowe and my friends Shamir Bhuiyan and Nur Azam for proofreading the thesis.

I am grateful to the University of Melbourne for the financial and logistic support throughout the tenure of my postgraduate research. I would like to thank the Head of the Department Prof. Justin Zobel and all the staff of the Faculty of IT for their support and encouragement.

I would also like to thank my friends Nahian, Muzammel, Rasel, Rezuwan, colleagues vii

(8)

Finally, I would like to thank my parents, parents-in-law, brothers, and sisters who have always encouraged me to pursue my higher studies. I am grateful to my lovely wife for her understanding and sacrifices during critical times of my studies. I deeply ac- knowledge the unconditional love, persistent encouragement and immeasurable support of my wife Sharmin Nahar and my daughter Maymunah Alhafizah.

viii

(9)

Preface

The thesis has nine chapters and the major chapters are 2, 3, 4, 5, 6 and 7.

Chapter 2 is based on the following publication:

Md Akter Hussain, Alauddin Bhuiyan and Ramamohanarao Kotagiri, Progress on Analysing OCT imaging on Retina and Choroid: A Review” (To be submitted).

Chapter 3 is based on the following publications:

Md Akter Hussain, Alauddin Bhuiyan, Andrew Turpin, Chi D Luu, R Theodore Smith, Robyn H Guymer and Ramamohanarao Kotagiri, ”Automatic Identification of Pathology Distorted Retinal Layer Boundaries Using SD-OCT Imaging.” IEEE Transac- tions on Biomedical Engineering 64.7 (2017): 1638-1649.

Md Akter Hussain, Alauddin Bhuiyan and Ramamohanarao Kotagiri. ”Retinal Cross Sectional Layer Segmentation using Optical Coherence Tomography.” 2nd Annual Doctoral Colloquium, The University of Melbourne. 2014.

Md Akter Hussain, Alauddin Bhuiyan, Hiroshi Ishikawa, R Theodore Smith, Joel S.

Schuman and Ramamohanarao Kotagiri. ”An Automated Method for Choroidal Thick- ness Measurement from Enhanced Depth Imaging Optical Coherence Tomography Images”, Computerized Medical Imaging and Graphics. ( Under review )

Kokroo, Aushim, Alauddin Bhuiyan, Md Akter Hussain, Ramamohanarao Kota- giri, Meleha Ahmad and Theodore Smith. ”Validation of an Automated Software for

ix

(10)

Md Akter Hussain, Alauddin Bhuiyan and Ramamohanarao Kotagiri. ”Automatic Detection of Choroid-Sclera Interface in EDI-OCT Images.” 4th Annual Doctoral Collo- quium, The University of Melbourne. 2016.

Md Akter Hussain, Alauddin Bhuiyan, Chi D. Luu, Robyn H. Guymer, Hiroshi Ishikawa, Joel S. Schuman and Ramamohanarao Kotagiri. ”A robust and reliable 3D segmentation Method for the retinal layers from Optical Coherence Tomography imaging”, Computer Methods and Programs in Biomedicine. ( Under review )

Md Akter Hussain, Alauddin Bhuiyan, Chi D. Luu, Robyn H. Guymer, Hiroshi Ishikawa, Joel S. Schuman and Ramamohanarao Kotagiri. ”Novel Automatic Ap- proach of Computing Eight Biomarkers for Retinal and Neuropathy Diseases in Macula and ONH Centred SD-OCT Imaging” ( To be submitted )

Md Akter Hussain, Alauddin Bhuiyan and Ramamohanarao Kotagiri. ”Disc segmentation and BMO-MRW measurement from SD-OCT image using graph search and tracing of three bench mark reference layers of retina.” Image Processing (ICIP), 2015 IEEE International Conference on. IEEE, 2015.

Md Akter Hussain, Alauddin Bhuiyan and Ramamohanarao Kotagiri. ”Automatic Retinal Minimum Distance Band (MDB) Computation from SD-OCT Images.” Digital Image Computing: Techniques and Applications (DICTA), 2015 International Conference on. IEEE, 2015.

Md Akter Hussain, Alauddin Bhuiyan, Chi D. Luu, Robyn H. Guymer, Hiroshi Ishikawa, R Theodore Smith, Joel S. Schuman and Ramamohanarao Kotagiri. ”Clas-

x

(11)

sification of Healthy and Diseased Retina Using SD-OCT Imaging and Random Forest Algorithm” ( To be submitted )

xi

(12)

AD ICG

AMD Age-related Macular Degeneration ILM Internal Limiting Membrane aprxONL the approximate locations of ONL INL Inner Nuclear Layer aprxRNFL the approximate locations of RNFL IPL Inner Plexiform Layer aprxRPE the approximate locations of RPE IZ Interdigitation Zone aprxTRL the approximate locations of Three

Reference Layers LBP Local Binary Pattern

AUC An Area Under the receiver operator

characteristics Curve MDB Minimum Distance Band

BM Bruch’s membrane MUE Mean Unsigned Error

BMO Bruch Membrane Opening MZ Myoid Zone

BMO- MRW

Bruch Membrane Opening Minimum

Rim Width nGA nascent Geographic Atrophy

BoW Bag-of-word OCT Optical Coherence Tomography

Cc Choriocapillaris OCV Outer Choroidal Vessel

CED Canny Edge Detection ONH Optic Nerve Head

CERA Centre for Eye Research Australia ONL Outer Nuclear Layer CFP Colour Fundus Photography OPL Outer Plexiform Layer CNV Choroidal Neovascularization OSL Outer Segment Layer

CSI Choroid Sclera Interface PCA Principle Component Analysis

CTh Choroidal Thickness PL Photoreceptor Layer

CV Choroidal Vessel PS-OCT Polarization-Sensitive Optical Coher-

ence Tomography

CWS Cotton Wool Spots RBC the complex of the RPE/ BM/ Chorio-

capillaris)

DC Dice Coefficient RMSE Root Mean Square Error

DIN Depth-based Intensity Normalization RNFL Retinal Nerve Fiber Layer

DME Diabetic Macular Edema ROI Region Of Interest

DR Diabetic retinopathy RPE Retinal pigment epithelium

EDI- OCT

Enhanced Depth Imaging Optical Co-

herence Tomography SA Simulated Annealing

ELM External Limiting Membrane Sch Suprachoroid

EZ Ellipsoid Zone SD-OCT Spectral Domain Optical Coherence

Tomography

FA Fluorescein Angiography SEAD Symptomatic Exudate-Associated De- rangement

GA Geographic Atrophy SLO Scanning Laser Ophthalmoscopy

GCL Ganglion Cell Layer SR Stable Reference

GT Ground Truth SS-OCT Swept Source Optical Coherence To-

mography

HRC Hyper-Reflective Complex TD-OCT Time Domain Optical Coherence To- mography

HRS Hyper-Reflective intra-retinal Spots TRL Three Reference Layers ICC Interclass Correlation Coefficient VMT vitreomacular traction

xii

(13)

List of Figures

1.1 A surface image of the retina using Colour Fundus Photography with outlining the major components/substructures of the retina. . . 3 1.2 A cross section of the retina and the choroid using Optical Coherence To-

mography (OCT) imaging, slice is known as B-scan. Outlining the layers of the retina and the choroid in the image. Details of these layers are discussed in Chapter 2 (section 2.2.1.4). . . 3 2.1 The human eye¹. . . 18 2.2 Colour Fundus Photography (CFP) Image of a portion of the retina where

optic disc, cup, macula, artery and vein are shown. The cross-section of the optic disc and macula (along a green line at the elongated CFP image) displayed at the right of the corresponding CFP image captured by SD- OCT technology. The 3D view of the optic disc and macula-centred images illustrated at the top and bottom are constructed by multiple SD-OCT scans. 19 2.3 The layers of the retina and the choroid, and the sclera. . . 21 2.4 Pathologies shown in CFP (left) and SD-OCT (right) images. (a) Cotton

wool spots / Soft Exudates¹; (b) Hard Exudate [1]; (c) drusen; (d) Geo- graphic Atrophy; (e) Hyper-Reflective intra-retinal Spots (HRS) (it is not visible in CFP) [2]; (f) Choroidal neovascularization (it is not visible in CFP) [3]. Arrow sign indicated the pathologies. . . 24 2.5 Examples of diseases’ effects on vision¹. . . 25 2.6 (a) SD-OCT B-Scans and (c) and (e) are Enface images of (b) and (d) CFP

images respectively. The atrophic regions (white) in the enface image is much clearer than CFP images [4]. The red lines showing the outer and inner segmentation lines (arrowheads in a) are used to generate the enface. 31 2.7 Example of BMO-MRW and MDB. . . 33 2.8 An example of the attenuation coefficient of the RNFL [5]. . . 34 3.1 A Colour Fundus Photography image showing the retinal surface (top left)

and a macula centre SD-OCT B-scan image (top right), a portion of the cross section across green line (top left image), defining the layers in the SD-OCT B-scan image. Proposed segmented boundaries are delineated in an SD-OCT image (bottom). . . 57

xvii

(18)

3.2 SD-OCT B-Scan images of the retina showing VMT, Drusen and OCT defined atrophy: (a) depicts drusen (red ellipse) and VMT (red arrow sign);

(b, c, and d) contain drusen, distortion of layers and morphological changes in each image; (e) contains OCT defined atrophy; the left red ellipse of (f) is OCT defined atrophy and layers loss and the right red ellipse of (f) is drusen and layer loss. . . 57 3.3 Flow diagram of the proposed method . . . 60 3.4 The TRL approximate detection: (a) is a smooth and a cropped portion of

the retinal SD-OCT B-Scan; (b) manually delineated TRL; and (c) automatically discovered aprxTRL by our proposed method (red, green and blue lines represent the RNFL, ONL and RPE layer locations respectively). . . . 62 3.5 (a) A B-scan image showing one A-scan as a green line; (b) the intensity

profile of the A-scan in (a) (green line on a); (c) is an A-scan from a different image, showing a different pattern of intensity profile; and (d) is used to fit to the actual signal for finding the aprxTRL. The red, green and blue circles define the RNFL, ONL and RPE layer positions respectively. The arrow sign on (c) indicates that pixels under RPE have higher intensity than the RNFL layer. . . 64 3.6 Flow diagram for identifying a boundary in our proposed method. . . 65 3.7 An example of MZ-EZ boundary detection steps: (a) the input SD-OCT

B-Scan image; (b) the edge Image after applying Canny edge detection; (c) the edge pixels having positive intensity gradient; (d) candidate pixels; (e) a magnified image of the red region of (d), each colour represents different pixels-groups, and black circles represent the end pixels and node of the graph; (f) An example of the fully connected graph representation of the boundary detection problem (s and e is two special node added automatically for defining the start and destination for the shortest path algorithm).

(g) corresponding pixel-groups obtained from the shortest path algorithm;

and (h) the MZ-EZ Boundary (yellow line). . . 67 3.8 An example of the computation of the weight for finding the slope discon-

tinuity (ψ) . . . . 69 3.9 An example of the computation of the weight for finding the layer’s non-

associativity (γ). . . . 70 3.10 The start and end nodes are specially and automatically handled in our

method. Only horizontal distances are considered to compute the edge weight for all nodes to start and end nodes. . . 71 3.11 Two examples of aprxTRL refinement (shown in smoothed SD-OCT im-

ages). (a) and (c) are the aprxTRL before refinement; (b) and (d) are the refined aprxTRL of (a) and (c) respectively. Red, green and blue lines represent the RNFL, ONL and RPE layer positions respectively. . . 74 3.12 The segmentation outputs by manual, OCTRIMA-3D, Chiu et al., Dufour

et al., Iowa Reference Algorithm, AURA Tools and our proposed method.

The ILM-RNFL, MZ-EZ, IZ-RPE and RBC boundaries are delineated using red, yellow, green and magenta lines respectively. . . 85

xviii

(19)

LIST OF FIGURES LIST OF FIGURES

4.1 Colour Fundus Photography image (left) and EDI-OCT image (middle) in a healthy eye. B-scan mode is demonstrated by the cross sectional image corresponding to the green line (left); A-scan mode is represented by red vertical line (right). . . 91 4.2 Flow diagram of the proposed method . . . 92 4.3 (a) The region of interest of the choroid, (b) The smooth image after inten-

sity normalisation and (c) Depth- based intensity normalised (DIN) image 94 4.4 Vessel pixels after: (a) clustering; (b) morphological operation; (c) morpho-

logical closed operation; and (d) applying dynamic distance filter. . . 96 4.5 (a) Shows black colour pixels found after clustering, red pixels in (b) and

(c) are selected as vessel after applying distance filter. . . 97 4.6 (a) Surface pixels, (b) following removal of top 100 µm surface pixels, (c)

probability map for the surface pixels, (d) candidate pixels (pixels (red) with maximum probability in each A-scan), (e) first order polynomial line on the candidate pixels (green) (f) following removal of candidate pixels (magenta). . . 98 4.7 (a) The DIN image, (b) approximate vessel, (c) approximate CSI (green

line) on the surface image (white colour), (d) approximate OCV boundary (red line), (e) approximated CSI (green line) and OCV boundary (red line) on the intensity normalised image. . . 99 4.8 (a) Original image (b) segmented output of ILM-RNFL (green line), RBC

(red line) and CSI (yellow line) (c) segmented output of choroidal vessels (magenta lines). . . 102 4.9 Dice Coefficient against dB of the volume between first manual grader and

automatic methods. . . 106 4.10 The output of our proposed method in the SS-OCT retinal scan presented

with pathologies. . . 108 5.1 Macula centred retinal image (a) Near infra-red image; (b) Volume or 3D

reconstruction of the retina from OCT scans; (c) A B-scan image (cross- section of the retina through the green line in Fig. a and b) and (d) Retinal Layers are delineated in a B-scan image. . . 112 5.2 The basic flow diagram of the 2DS algorithm for the boundaries of the

retinal layers. . . 118 5.3 The basic flow diagram of our proposed 3D segmentation. . . 128 5.4 Segmentation result by state-of-the-art and our proposed 3D automatic

methods on normal subject. . . 133 5.5 Segmentation result by state-of-the-art and our proposed 3D automatic

methods on a subject with AMD. . . 133 xix

(20)

5.6 Some examples of detection in the presence of different pathologies using our proposed method. The CSI is not detected in (e) and (f) due to not captured properly. (a) and (b) In the presence of large Cyst (red arrow signs) of DME patient (c) In the presence of small Cyst (red arrow signs) of DME patient;; (d) In the presence of small Cyst (red arrow sign) and lesion at the inner retina (blue arrow signs) of DME patient; (e) and (f) In the presence of vitreomacular traction (green arrow sign), drusen (pink arrow sign) and RPE detachment (yellow arrow sign) of AMD patients. In some places of the image (f) have lost the ONL layer properties (orange arrow sign) as well. (a, b, c, and d) Images are collected from a public dataset (DUKE university) [1]. . . 134 5.7 A pictorial example of our proposed method segmentation. (a) 3D ren-

der of a volume; (b) A 3D render image of RNFL (blue), ONL (magenta) and RPE (yellow) Layers; and (c) A 3D render image of RNFL (blue), and Choroid (green). . . 135 6.1 Macula centred retinal image. (a) SLO image; (b) A 3D or volume of the

retina from OCT image; (c) A B-scan image (cross-section of the retina through the green and yellow lines in Fig. (a) and (b) respectively) and (d) Retinal Layers are delineated in a B-scan image (e) Optic disc centred Enface image (Cup: red area; Rim: green area) (f) A B-scan of Optic Disc centred Retina (g) A B-scan showing drusen (blue) (h) A B-scan showing Hyper-Reflective intra-retinal Spots (HRS) (pointed by red arrow). . . 142 6.2 ILM-RNFL boundary detection in the presence of ONH. (a) an ONH-centred

B-scan; (b) positive gradient edge pixels; (c) Candidate pixels for the ILM- RNFL boundary; (d) potential ONH region (red colour A-scans); (e) mod- ified candidate pixels in the potential ONH region after applying more smoothing operations during edge detection; (f) The ILM-RNFL boundary at the ONH-centred SD-OCT B-scans the red line. . . 146 6.3 ONH detection. (a) Enface image of a portion of the ONH centred retina;

(b) A cross section of the retina (B-scan, along green line in (a)) with delineating ILM-RNFL (red line); (c) Approximate RNFL layer (pink)and approximate three layers positions (RNFL: red line, ONL: green line, RPE:

blue line) (d) Potential ONH regions: green line for distance between polynomial line and ILM-RNFL boundary, yellow line for the pattern of the un- like position of the approximate ONL and RPE, blue line for the intensity disorder between the approximate ONL and RPE; (e) Detected initial positions of the ONH (pink) and the best circle (red) fit; (f) ONH boundary(red vertical line), ILM-RNFL (top continuous red line) and BM boundary (bottom red lines at the left and right of the ONH). . . 147 6.4 Automated segmentation output. Segmented (a) macula centred normal

subject, (b) AMD subject in the presence of drusen and (c) ONH centred glaucoma subjects. . . 149

xx

(21)

LIST OF FIGURES LIST OF FIGURES

6.5 Quantification output of layer thickness (Biomarker-1). (a) Manual and (b) automatic thickness map of the macula centred retina from a subject of the AMD dataset. . . 151 6.6 Hyper-Reflective intra-retinal Spots (HRS) in the retinal SD-OCT (Biomarker-

2). (a) SD-OCT B-scan (b) manual ILM-RNFL boundary (red line) and HRSs (green) (c) automatically detected ILM-RNFL boundary (red line) and HRSs (green). . . 152 6.7 2D and 3D drusen visualisation (Biomarker-3). (a) Manual and (b) auto-

matic detection of drusen, blue in the B-scan (top) and 3D view of drusen (bottom). . . 153 6.8 Cup and Rim in a SD-OCT B-scan [6]. BMO points (green dots) indicate

the disc area. The reference plane (red line) was set above the base plane (BMO plane, green line) at a height 120 µm. Intersections of the ILM-RNFL and the reference plane indicated the cup area (green dot). . . 155 6.9 A retinal B-scan in optic disc region with BMO-MRW and MDB. . . 156 7.1 The flow diagram of the proposed classification method. . . 164 7.2 (a) A SD-OCT B-scan (b) manual ILM-RNFL boundary (red line) and HRS

(green) (c) automatically detected ILM-RNFL boundary (red line) and HRS (green). . . 165 7.3 (a) A SD-OCT B-scan with delineating drusen by the blue colour (b) Drusen

in 3D view of an SD-OCT volume of an AMD patient. . . 166 7.4 The curviness of different MZ-OZ boundaries (red) with a different curve

using our proposed method. RBC boundary is in green. . . 168 7.5 (a) A 3D render image of the retina with the choroid constructed from an

SD-OCT volume; (b) Segmented layers of the retina and choroid; (c) The complex of the EZ, IZ, and RPE in a different colour in the gray-scale retinal SD-OCT image. . . 170 8.1 The graphical user interface of the developed system using proposed method.178 8.2 The flow diagram of the OCTInspector System. . . 178 8.3 The correction mode where Choroid-Sclera Interface (CSI) is selected for

correction. . . 179 8.4 (a) 3D View of a segmented macula centred OCT image; (b) 3D view of a

segmented RNFL and Choroid in macula centred OCT image; (c) 3D view of a segmented ILM surface and ONH; (d) 3D view of the RPE in ONH centred OCT image; (e) Enface image of the RNFL, GCL, IPL from Mac- ula Centred OCT Image; (f) Enface image of the RPE Layer from macula- centred OCT image that shows the drusen by brighter intensity; and (g) 3D view of drusen from a macula-centred SD-OCT volume. . . 180 9.1 A brief summary of the thesis contribution and the future research direction.184

xxi

(22)

(23)

List of Tables

1 List of abbreviations in lexicographical order. . . xii 1.1 Association of the biomarkers with eye diseases. . . 4 2.1 Significant features of the imaging technology of the retina and the choroid

based on [7–11]. . . 28 2.2 The noise reduction methods on OCT images. . . 37 2.3 The segmentation methods for various structures of the retina and the

choroid from OCT images. . . 39 3.1 Parameter selection using the Simulated Annealing (SA). . . 78 3.2 The mean ±standard deviation of the RMSE in pixels for all boundaries

on the Chiu et al. public dataset. . . 81 3.3 The mean ±standard deviation of the RMSE in pixels for all boundaries

for Tian et al. (OCTRIMA-3D) data set of normal eyes. . . 82 3.4 The mean±standard deviation of the RMSE in pixels for various methods

on the CERA data set. . . 84 4.1 The performance of the boundary of RBC and CSI, Choroid and CTh of the

Tian et al., Chen et al. and our proposed methods. (units in pixel) . . . 107 5.1 The parameters for detecting the boundaries of the retinal tissue layers. . . 120 5.2 The mean and standard deviation in pixels of the RMSE for boundary po-

sition on normal subject’s dataset. . . 131 5.3 The mean and standard deviation in pixels of the RMSE for boundary po-

sition on AMD subject’s dataset. . . 132 5.4 The means and standard deviation of the evaluation protocols for CSI po-

sitions and the thickness of the choroid. . . 132 5.5 The mean and standard deviation in pixels of the RMSE for boundary po-

sition on AMD subject’s dataset for different resolution (number of B-scans per 6µm) . . . 132 6.1 The association of the biomarkers with DME, AMD and glaucoma. . . 143 6.2 The mean percentage of absolute error between manual and automatic de-

tection of layers thickness (Biomarker-1). . . 151 6.3 The summary of performance evaluation of the Biomarker 2 to 7. . . 159

xxiii

(24)

7.1 Performance of four state-of-the-art and proposed methods on partial DUKE dataset (D-1) considering only normal and DME patients (because Ven- huizen et al., Lemaitre et al., and Sidibe et al. used to classify only them). . 172 7.2 Average of 10 confusion matrixes on 15-fold cross-validation test for the

proposed classification model using Random Forest. . . 173 7.3 The accuracy for different machine-learning algorithms for the classifica-

tion model based on the proposed features. . . 174

xxiv

(25)

Chapter 1

Introduction

Life expectancy is increasing and vision needs to be preserved to maintain a good quality of life for individuals. Consequently, vision loss is an alarming issue for the ageing population. The top three common eye diseases which result in irreversible vision loss are Diabetic Macular Edema (DME), Age-related Macular Degeneration (AMD) and glaucoma. These diseases often remain undiagnosed or are only diagnosed late and can cause permanent loss of vision which is impossible to reverse [7, 8, 12, 13]. DME mainly affects diabetic patients and the prevalent cases of DME are expected to grow to over 300 million (globally) within the next few years [14]. AMD and glaucoma affect mainly aged people. AMD affects nearly 35% of adults who are over 80 years of age; glaucoma accounts for 9-12% of all cases of blindness. The number of AMD and glaucoma patients are expected to increase by approximately 150% over the next few years due to an increase in the ageing population [7, 13]. As a consequence, a large proportion of the world’s health budget needs to be spent on screening, diagnosis and treatment of these diseases. The costs for individuals suffering from such diseases can be enormous. There- fore, early detection and treatment of those diseases can save vision, money and provide a better quality of life for individuals.

Our visual system involves both the eyes and brain. The eye has many parts or structures such as lens, retina, choroid, sclera. The light comes through the lens of the eye and falls onto the retina which converts the light into an electrical signal for passing to the brain that processes the signal to make sense to us. The choroid is located under the retina and its main purposes are to provide oxygen, nutrition to the retina and to absorb

1

(26)

excess light to protect the retina. The retina has many substructures such as optic disc, blood vessels (artery and vein), macula as shown in colour fundus image, Fig. 1.1. Colour fundus image can only show the surface of the retina and can provide a detailed image of the retina substructures’ surface. The retina has ten different layers of tissue as shown in Fig. 1.2. The functions of these layers are discussed in Chapter 2 (Section 2.2.1.4).

The cross sectional view of these tissue layers can be observed using Optical Coherence Tomography (OCT) imaging: each slice is known as a B-scan (see Fig 1.2) [7]. High resolution and the speed of the OCT technology also allow constructing a 3-Dimensional (3D) view of the retina by capturing and combining multiple OCT images. Since the choroid is located under the retina, it is not as easy to observe from the outside as the retina.

However, advances in OCT imaging technology allow capturing the cross-section of not only retinal tissue layers but also the choroid as shown in Fig. 1.2. OCT imaging can cap- ture the retina and the choroid structures in detail and can extract biomarkers from these structures for understanding and monitoring the progression of eye diseases. Ophthal- mologists have found some morphological changes such as variation in layers’ thickness in the retina and the choroid and the presence of cysts (a risk factor for DME), and drusen (a risk factor for AMD) due to these eye diseases before there is any noticeable deteriora- tion in vision experienced by the individual [15–18]. The morphological changes noticed for eye diseases are listed in Table 1.1.

Research studies associated with understanding the progression of the diseases required analyses of large numbers of OCT images and quantifying many potential biomarkers. Traditional methods of quantifying biomarkers involving humans are no longer fea- sible or cost-effective on large-scale datasets as high-resolution images from OCT imaging create a huge volume of images. For example, it is possible to collect up to 400 OCT images per person for both eyes every 3 months [24]. Moreover, human grading is prone to inaccuracy, more grading variability and subjective bias. On the other hand, automatic grading will allow more consistent and effective measurements on a very large scale and consequently may give an opportunity to gain a new reliable insight into many eye diseases [7, 8]. This has given impetus to building automatic tools for segmentation of the retinal layer from OCT images since 1995. Some segmentation methods have been pro-

(27)

3

Figure 1.1: A surface image of the retina using Colour Fundus Photography with outlining the major components/substructures of the retina.

Figure 1.2: A cross section of the retina and the choroid using Optical Coherence Tomog- raphy (OCT) imaging, slice is known as B-scan. Outlining the layers of the retina and the choroid in the image. Details of these layers are discussed in Chapter 2 (section 2.2.1.4).

posed to enhance the clinical benefit of the OCT such as detecting and quantifying the pathologies and layer thicknesses. However, none of the methods are capable of segmenting layers with good accuracy in all conditions due to different challenges found in

(28)

Table 1.1: Association of the biomarkers with eye diseases.

Biomarkers Association of the biomarkers with the eye disease

Layer Thickness

RNFL, GCL, and IPL layers of the retina were significantly thin- ner in eyes with glaucoma - approximately 20% less width than age-matched normal eyes (p < 0.001) [17]. Mean retinal thickness was reduced significantly in early AMD patients, approximately 10%

(p=0.008) than age-matched normal eyes [19].

Quantification value of Hyper- Reflective intra- retinal Spots (HRS)

The presence of HRS is a characteristic finding of the various stages of DME and is a key risk factor for the development of more advanced stages of DME [18].

Quantification value of Drusen

The presence of large macular drusen ( 125 µm) is a characteristic find- ing of the early stages of AMD and is a key risk factor for the development of more advanced stages [20].

Cup-disc ratio A cup-disc ratio more than 0.5 is a risk indicator of glaucoma [15].

Bruch’s membrane opening - minimum rim

width (BMO-

MRW)

The visual sensitivity of glaucoma patients is significantly correlated to the BMO-MRW (r = 0.32, p < 0.001), which has a higher correlation than RNFL Thickness [21]. BMO-MRW in normal patients is 307± 84.3 µm whereas early glaucoma patients have 211± _{60.5 µm} [22].

Minimum dis-

tance Band

(MDB)

The correlation coefficient between the MDB and cup-disc ratio are

−0.88 and−0.56 for MDB value and area respectively with p < 0.05 which means MDB is highly correlated to glaucoma like cup-disc ratio [12].

Attenuation Co- efficient of the RNFL

The severity label of glaucoma increases with decreasing the RNFL’s attenuation coefficient [23]. Schoot et al. [23] found a significant structure-function relationship between the attenuation coefficient and the visual field’s mean defect.

Reflectivity value

of EZ layer Early AMD patients have an average value of 1.73, and the control patients have an average value of 2.27 [16].

RNFL: Retinal Nerve Fiber Layer; GCL: Ganglion Cell Layer; IPL: Inner Plexiform Layer; EZ: Ellipsoid Zone

OCT images. Such challenges include: unpredictable changes in the retina and choroid due to pathologies and anatomical structures such as optic disc, macula; inconsistent contrasts in the homogeneous areas due to noise and imaging technology. Considering these essential requirements, in this study, an automatic 3D segmentation method for the retinal layers and the choroid from OCT images is proposed. Since OCT technology can show morphological changes of the retina and the choroid due to eye diseases, it is possible to design a classification model of eye diseases and its severity using this morphologi-

(29)

1.1 Research objectives 5

cal information. There are some methods proposed for the classification of diseases based on texture analysis of OCT images [1,25]. There is only one method available for the classification of diseases from clinically defined morphological changes, where biomarkers are extracted manually and results show excellent accuracy [26]. Therefore, in this study, novel methods have been developed for detecting pathologies such as drusen and extracting biomarkers such as changes of the layer thickness. Using these clinically derived biomarkers automatically, a classification model has been designed for identifying eye diseases that showed excellent accuracy.

1.1 Research objectives

This research aims to develop an automatic segmentation algorithm for the retinal layers and the choroid which can be utilised to quantify clinically derived biomarkers. Further- more, this project aims to quantify pathologies, such as inner retinal lesion (e.g. HRS) and outer retinal lesion (e.g. drusen). In particular, this research addresses the following problems: 1) automating 3D segmentation of the retinal layer and the choroid; 2) quantifying biomarker related to DME, AMD, and glaucoma, and 3) developing a retinal OCT-based classification model for classifying the eye diseases into DME, AMD and normal.

1.1.1 Automated 3D Segmentation algorithm for the retinal layer and the choroid

The main goal of this study is to develop an automatic method for segmenting the retinal layers and the choroid from OCT images. An automated method will help ophthalmologists to conduct large-scale early detection and to monitor eye diseases. Manual segmentation of the retinal layers and the choroid is very expensive and time-consuming because OCT technology provides high-resolution images with a large number of slices (B-scans).

Just one individual may have 400 OCT images for both eyes. Moreover, the quality of the manual segmentation is prone to inaccuracy, higher grading variability, and subjective bias and fatigue. Therefore, it is highly desirable to develop an accurate and robust automatic method for the segmentation of the retinal layers, the choroid and pathologies

(30)

of the retina as well as to provide their quantified values. However, that is a challenging task due to the uneven anatomical structure of the retina, unpredictable changes due to pathologies and inconsistent contrasts in homogeneous areas because of noise [8]. In the last two decades, a large number of studies have been done to develop an automatic tool to resolve this problem. However, none of the methods has performed well for the retinal layer segmentation as the accuracy of these methods is highly dependent on dataset due to the presence of various pathologies. Therefore, we need to develop an improved automatic segmentation method for retinal layers and choroid that can work reliably under various disease conditions and eye anatomical structures.

OCT images are contaminated by additive and speckle noises [27]. As a result, a homogeneous area can show uneven contrasts. Moreover, the intensity values decrease deterministically with growing imaging depth across layers due to the absorption and scattering of light in the retina and the choroid tissues which make detecting the choroid a challenging task. The structure of the retina is also different in the different regions of the retina such as optic disc and macula, as shown in Fig. 1.1 and also becomes more unpredictable in the presence of the pathologies such as geographic atrophy (GA) and drusen [8]. The presence of varying blood vessel sizes makes an unpredictable pattern in the Optic Nerve Head (ONH) region, another name of optic disc [8]. These properties make it difficult to segment ONH region. Additionally, the choroid differs largely from the retina, where large blood vessels and the structure of the choroid make an uneven pattern throughout the region. Thus, developing a robust method for segmenting the retinal layer and the choroid to handle those difficulties is a big challenge. Therefore, an automatic and robust method is required for handling the noise, intensity inconsis- tencies, different anatomical regions (the presence/absence of the ONH and/or macula), unpredictable changes due to the presence of pathologies, etc. to find accurate positions of the retinal layers and the choroid.

1.1.2 Biomarkers quantification

The goal of the retinal layer automatic segmentation is to extract biomarkers for the prediction of eye diseases. Such an automated system would enable large scale studies. It

(31)

1.2 Thesis contributions 7

can also help to establish new biomarkers by analysing their correlation with eye diseases. To extract reliable biomarkers, a segmentation method should be able to detect the pathologies with high accuracy. That detection is very challenging due to unpredictable positions, shapes, and sizes of the pathologies. Biomarker extraction also involves all ten retinal layers that need to be detected properly. The retinal layers’ detection in the diseased eye is more challenging due to unpredictable thickness variations in the layers and loss of the retinal layers. Consequently, thin layers become more challenging to detect accurately. Therefore, this thesis investigates the development of robust algorithms for accurate extraction of biomarkers.

1.1.3 Development of classification model for eye diseases

Prediction of eye diseases from OCT images is an important problem. It can help early detection of eye diseases and be able to calibrate the severity of the disease. Consequently, many research schemes have been suggested for designing a classification model. Most methods classify the diseases based on the texture of the OCT images. Since textures could misguide the model due to changes of the dataset and the presence of multi-type pathologies and noise, they are prone to error and thus a model based on the clinically- driven parameters is required [28]. To produce a robust classification method for an eye disease, we need to address the following: (a) accurate measurement of biomarkers; (b) selection of appropriate machine-learning algorithm, and (c) computation of the optimal hyper-parameters for the chosen machine-learning algorithms.

1.2 Thesis contributions

The contributions of this thesis are: a) To develop an efficient and accurate automatic method of segmenting all ten retinal layers and the choroid; b) Quantifying the biomarkers including segmenting the pathologies; and c) Classification of eye diseases; with addressing three important research problems as mentioned in the previous section.

In addition to the above contributions, we have developed a fully automated software system for OCT image segmentation, biomarkers extraction and classification. The

(32)

system is designed to manually correct any part of the automated segmentation. This functionality helps the system to continually improve its accuracy; corrected images can be used in the learning process. The details of each contribution are presented in the following subsections.

1.2.1 Automated 3D Segmentation algorithm for the retinal layer and the choroid

A robust and effective automatic method has been proposed for segmentation of the retinal layers and the choroid from the Spectral Domain Optical Coherence Tomography (SD-OCT) images or higher-resolution OCT images such as Enhanced Depth Imaging OCT (EDI-OCT) and Swept Source OCT (SS-OCT). The proposed method seamlessly works in the presence or absence of the ONH and/or macula. The proposed method also works in the presence or absence of pathologies and morphological changes due to disease. The segmentation of boundaries is achieved by modelling the problem as the shortest path graph problem. The edge pixels found from a Canny Edge Detection Algo- rithm form the nodes. The slope similarity to a reference line and node’s non-associativity (pixels not satisfying associated layer property) to the layer along spatial distances of the nodes are used for the computation of the graph edge weight. The edge weight is computed by addressing pathologies, macula and ONH related structural change of the retina. Since the choroid is considerably different from the retina, where large blood vessels and the structure of the choroid make the region uneven in the distribution of intensity, the approach of detecting retinal layers is not suitable for the Choroid-Sclera Interface (CSI) - the outer boundary of the choroid. A novel method has been proposed to normalise the intensity of the choroid region for providing an even distribution of the background. Finally, the CSI boundary is detected using a similar segmentation algorithm that is used for retinal layers; however, edge weights incorporate the effect of imaging technology and the anatomy of the choroid. Moreover, the tissues of the retina and the choroid are continuous in adjacent B-scans when the distance between the adjacent scans are very close. Therefore, very small changes of the boundaries from one B-scan to an adjacent B-scan are expected. This information helps to obtain correct boundaries where 2D automatic segmentation fails due to the presence of noise or various tissue

(33)

1.2 Thesis contributions 9

structures or pathologies. The method first detects the boundaries sequentially in the order of high contrast and the maximum gradient intensity to low contrast and minimum gradient intensity of the boundaries. This approach helps detect the low-contrast boundaries in a small Region of Interest (ROI), since ROI is defined using the already detected boundaries and adjacent B-scans. The reduction of the ROI helps to improve the accuracy and efficiency of the detection even in the presence of pathologies. Due to the differences among the patterns of ONH boundaries, the top boundary of the retina is detected by utilising approximate ONH region in the edge weight. Following this, ONH is detected by using the top boundary and three patterns (the absence of layers, dissimi- lar layer positions, and intensity pattern) of the ONH in the SD-OCT image. This method is evaluated using five datasets from four sources including two public datasets, which consists of 56 subjects where 55 are macula-centred volumes and one is ONH-centred SD- OCT. In those datasets, 36 subjects are AMD and glaucoma patients, and 20 subjects are healthy. Three different graders trace the boundaries for different datasets that serve as a gold standard for the automatic segmentation evaluation. In total, eight state-of-the-art methods (six methods are for retinal layer segmentation while the others are for the CSI segmentation) and the proposed method has been used to compare the accuracy of the automatic methods. The proposed method has also been shown to outperform the other eight state-of-the-art methods on every dataset.

1.2.2 Biomarkers quantification

In this study, a novel framework has also been proposed for measuring biomarkers that are already defined by ophthalmologists for retinal structures using SD-OCT images. A total of eight biomarkers are of interest to ophthalmologists. These are: retinal structural thickness; three morphological parameters of ONH; the volume of the pathologically al- tered tissue (lesion of the inner and outer retina); the relative intensity of the Ellipsoid Zone, and attenuation coefficients of the Retinal Nerve Fibre Layer. These are very significant for early screening of glaucoma, DME and AMD. The proposed 3D segmentation method delivers its primary goal of detecting the layers in any circumstances - for example, in the presence of pathologies or distorted layers due to disease. Furthermore, an

(34)

automated method for detecting pathologies (such as drusen and HRS) using the thickness of the layers and intensity profiling has been developed. The same protocol has been followed as defined by the previous studies that introduced these biomarkers when computing them on SD-OCT images that resulted in low margins of error between manual and automatic grading of the biomarkers. Two AMD patients (having a total of 98 B-scans) have been used for the evaluation of the quantified values of the biomarkers;

one glaucoma patient (total 200 B-scans) and one DME patient (total 97 B-scans). The automatic method achieved F1-scores of 0.79 and 0.70 for a HRS of the inner retina and drusen respectively, using manual grading as a gold standard. The mean error of the biomarkers’ quantified value is as low as 0.06.

1.2.3 Development of classification model for eye diseases

In this study, the first automatic method of clinically-derived features-based classification method of eye diseased patients from SD-OCT images has been proposed. The patients are classified into DME, AMD and normal. Total ten features have been used for developing the model of eye-disease classification. Ten features are comprised with the thickness of the retina and retinal layers, the volume of the pathologies such as drusen and HRS, curviness of the boundaries of the retinal layers. The classification model is then designed based on Random Forest. Experimental results with two datasets of 45 (a public dataset, 15 DME, 15 AMD and 15 normal) and 72 (combining the public and local datasets, 15 DME, 28 AMD and 29 normal) show the SD-OCT volumes have very good classification accuracy. The proposed method has achieved a high level of accuracy compared to the existing four state-of-the-art methods.

1.2.4 OCTInspector: A Complete Automated System for OCT Image Analysis

Although the proposed automatic segmentation shows excellent performance in segmenting the retinal layers and choroid, it can occasionally fail in some places and require manual correction. Therefore, we have provided a very efficient manual editing facility.

This uses the same graph representation as the automatic method but reduces the ROI

(35)

1.3 Thesis organisation 11

by the use of mouse pointer and clicks. The manual correction is made consistent by the use of nearby edge-pixels instead of the exact position of the user’s clicks. The system also provides an option to visualise any set of combination of the layers, sub-structures and pathologies in 3D as well as in enface image (frontal sections of retinal OCT scans, also called C-scan OCT). It also computes selected biomarkers that have been proposed in this thesis.

1.3 Thesis organisation

The rest of this thesis is organised as follows.

Chapter 2: Background

This chapter presents the background knowledge of the biological properties of the retina and the choroid. Following that, the major pathologies and disease manifestation in the retina and the choroid are discussed. The imaging technologies of the retina and the choroid are briefly presented. We also discuss various OCT technologies and imaging procedures. Finally, a literature review on the existing computer-aided segmentation methods for the retinal layers, the choroid, and pathologies; and classification model based on SD-OCT images are presented.

Chapter 3: 2D Segmentation (2DS) Algorithm for the Detection of Retinal Layers This chapter presents an automatic 2D segmentation (2DS) algorithm of the four significant boundaries of the retina that are distorted in the presence of pathologies (such as drusen) for the macula-centred SD-OCT image. This is the initial work towards developing the proposed 3D segmentation method. This chapter proposes a noise-reduction approach using anisotropic diffusion and Weiner filter. Then, an algorithm for finding the three reference layer positions approximately using prior knowledge of intensity and position of them has been proposed. Following that, the procedure of the boundary detection by constructing graph is discussed. Finally, performance evaluation and com- parison between automated methods is presented.

Chapter 4: 2D Segmentation (2DS) Algorithm for the Detection of the Choroid- Sclera Interface

(36)

In this chapter, an automated 2D segmentation (2DS) algorithm of choroid segmentation from EDI-OCT images is presented. A novel intensity-normalisation technique that is based on the depth of the choroid is used to equalise the intensity of all non-vessel pixels in the choroid region. Extension of the 2DS algorithm developed in Chapter 3 for the CSI is described. This method is tested on 190 B-scans of 10 subjects against manual segmentation by two expert graders and two state-of-the-art automated methods.

Chapter 5: 3D Segmentation (3DS) Algorithm for the detection of the retinal layers and the Choroid-Sclera Interface

In this chapter, an automated 3D segmentation (3DS) method of the retinal layers and the choroid from OCT images is presented. The previously- proposed 2D segmentation methods have been utilised in the 3D segmentation method. The selection of small layer-specific regions of interest from adjacent B-scans makes the method efficient, accurate and robust. The method on 250 B-scans from 10 (8 normal and 2 AMD) subjects has been evaluated by comparing the boundary positions and layer thicknesses marked by one grader. The performance of the proposed method has been compared with five state-of-the-art methods and the proposed method showed a significant improvement in accuracy.

Chapter 6: The Optic Nerve Head Detection and Eight Prominent Biomarkers Ex- traction

In this chapter, a proposal is made for a unified method of segmenting ONH from SD-OCT images that allow segmenting the retinal layers seamlessly in ONH or other regions. Furthermore, an automatic method of quantifying retinal biomarkers after segmenting pathologies, such as drusen and HRS, is proposed. Finally, eight clinically-useful biomarkers of the retinal diseases are computed automatically using a defined protocol by ophthalmologists. A low margin of error between manual and automatic grading of the biomarkers is found.

Chapter 7: Classification model of Diseased patients

In this chapter, an automatic method of classification of SD-OCT images for identification of patients into three different classes (DME, AMD and Normal) is proposed.

This classification model is designed based on ten features using Random Forest. Exper-

(37)

1.3 Thesis organisation 13

imental results have been used with two datasets including a public dataset. Classifi- cation of three classes (DME, AMD and normal) and two classes (diseased and normal) is performed and compared with state-of-the-art methods and using many classification methods.

Chapter 8: OCTInspector: A Complete Automated System for OCT Image Analysis In this chapter, the features and functionality of the OCTInspector are presented.

Chapter 9: Conclusion and Future Research Direction

This chapter summarises the contributions of this thesis and discusses possible areas for future research.

(38)

(39)

Chapter 2

Background

This chapter presents background knowledge of this research including the anatomy, the imaging technology and ending with identifying the research work done on the Optical Coherence Tomography (OCT) imaging of the retina and the choroid. This chapter is based on the following publication:

Md Akter Hussain, Alauddin Bhuiyan and Ramamohanarao Kotagiri. ”Progress on Analysing OCT imaging on Retina and Choroid: A Review”. (To be submitted).

2.1 Introduction

The eye is the main organ of the human visual system. The light enters the eye, the retina converts it into an electric signal and sends it to the brain for us to make sense [8]. The retina is thus called an extension of the brain. The choroid protects the retina from the harm of excess light as well as supplying nutrition [29]. The retina and the choroid are two structures of the eye and constructed by different tissues. They also have different substructures, such as an optic disc in the retina that contains nerve cells which trans- fer the electric signals to the brain. Consequently, many eye diseases and other systemic diseases such as strokes manifest themselves in the retina, the choroid and their substructures [30]. Eye diseases include ocular diseases, such as macular degeneration and glaucoma, which are the most significant causes of blindness in the developed world [7, 13].

Systemic diseases include diabetic retinopathy from diabetes, hypertensive retinopathy from cardiovascular disease, and multiple sclerosis [14]. Consequently, for the past few decades, ophthalmologists prefer to diagnose eye diseases by investigating the retinal

15

(40)

tissue structure and the choroid. With proper techniques, the retina is visible through the pupil and is accessible non-invasively for imaging. Since the choroid is behind the retina, it is not as easy as the retina to visualise from the outside. However, advances in imaging technology (e.g. spectral domain optical coherence tomography) permit them to be viewed with high resolution. Therefore, ophthalmologists are keenly interested in image-based diagnoses for the eye diseases due to accuracy and the ability to monitor disease progression.

Many non-invasive techniques are available for two-dimensional (2D) imaging of the retina such as Colour Fundus Photography (CFP) and InfraRed (IR). [7, 8]. More invasive techniques such as Fluorescein Angiography (FA) and IndoCyanine Green (ICG) angiography that require dye injection, are used for imaging functional retina and choroid.

However, advance-imaging technology, Spectral Domain Optical Coherence Tomogra- phy (SD-OCT) allows non-invasive procedures to obtain proper 3-Dimensional (3D) images of the retinal tissue structure. Enhanced depth imaging mode of SD-OCT technology can image choroid with high resolution [31]. Using biomarkers, for example, are where tissues or pathologies from the images are used for diagnosis of eye diseases and their progression. Extracting biomarkers requires segmentation of the OCT image for identifying tissue structures of the retina, choroid and optic disc. Segmentation can be performed either manually or automatically. Manual segmentation is generally trust- worthy due to expert human grading but it is time-consuming, expensive and can be unreliable for large-scale imaging [7, 8]. The problem becomes even more difficult with high-resolution images. Also, manual segmentation is prone to inaccuracy, more grading variability and subjective bias of the graders. As a result, manual segmentation is infeasible in practice. These limitations demand automatic segmentation methods. Therefore, many attempts have been made for segmenting the various structures of the retina and the choroid [7,8,25]. Automatic segmentation is very cheap and fast and can be employed for diagnosis and monitoring of progression of eye diseases. However, segmentation of OCT images is a challenging task due to the presence of noise, pathologies and different anatomical structures captured in the image. The noise in the images makes for uneven distribution of the intensity of retinal tissue structures. Diseased eyes have different

(41)

2.2 Eye, Retina and Choroid 17

types of pathologies that severely affect the tissue structures. Therefore, researchers have found the segmentation of the structures is very challenging. Some automatic methods succeeded in segmenting accurately high quality and fine image but performed poorly in the diseased eyes where accurate segmentation is most desired. In this chapter, we will review automatic segmentation of different structures and sub-structures of the retina and the choroid. In a nutshell, this review will cover the following topics.

• Diseases manifestation in the retina and the choroid

• Imaging based Diagnosis method of the retina and the choroid

• Biomarkers of the retina and the choroid including pathologies from OCT image

• Image analysis techniques based on OCT technology for:

– Detection of retinal and choroidal layers;

– Detection of pathologies;

– Extraction of biomarkers;

– Prediction of eye diseases using classification models.

2.2 Eye, Retina and Choroid

The eye and the brain constitute the human vision system. Figure 2.1 is a cross-section through the eye showing its major structures [32]. A ray of light passes through the cornea, the anterior chamber, the pupil, the lens and the vitreous, and is then focused on the retina [8]. The choroid, the sclera and the retinal pigment epithelium of the retina absorb any excess light and thus protect the eye from harmful light. The light is focused on 125 million receptors, called rods and cones, in the photoreceptor layer of the retina [33]. These receptors are nerve cells and, when the light hits the receptors, they emit electrical signals which are passed to the brain. The brain processes the signal and makes sense of the image. In the following subsection, we will look at the retina and the choroid in detail.

(42)

Figure 2.1: The human eye¹.

1http://webvision.med.utah.edu/

2.2.1 The retina

The retina is the inner part of the eye and a central entity of vision [33]. The major components of the retina are optic disc, macula and blood vessels as shown in Fig. 2.2. Different types of tissue cells comprise the retina, which is divided into ten layers [34, 35]. The following section elaborates on these components and the layers of the retina.

2.2.1.1 The optic disc

The optic disc is called a blind spot because it has no light-sensitive rods or cones tissues.

It is also known as Optic Nerve Head (ONH) [15]. In the CFP image, the optic disc is brightest in intensity as shown in Fig. 2.2. In the OCT image, the optic disc region is shown as a fall in layers due to the absence of retinal tissues. In the optic disc, there is a pink neuro-retinal rim containing the nerve fibres and a central pale area (cup) devoid of nerve fibres, shown in Fig. 2.2 (red circle and cup enclosed rim is by the green circle).

Automatic Optical Coherence Tomography Imaging Analysis for Retinal Disease Screening